{"id":"9e81dd8f-4ec3-4438-9252-351208b577ba","arxiv_id":"2411.10945","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A new egocentric 360-degree video dataset and a weakly-supervised model, FDPN, detect abnormal events and their directions around visually impaired wearers, reporting top scores on VIEW360, UCF-Crime, and Shanghaitech.","lead":"This paper introduces a new dataset of 360-degree videos, filmed with a neck-worn camera, showing threats like pickpocketing that people with visual impairments might face. It also presents a model that spots these unusual events and points out which direction they come from.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The frame-level training signal in Eq. (4) provides no within-snippet supervision, so the claimed boundary-level improvement is not entailed by the objective and is not isolated by any ablation.","rationale":"The reader's weakest assumption correctly identifies the pseudo-label supervision as the point where the method's load-bearing premise fails to follow from the equations. The central claim has two parts: FDPN predicts at frame level and achieves state-of-the-art results. The SOTA comparisons are also imperfect, with per-dataset teacher selection, no error bars, and no code or data release, but those are evidentiary weaknesses that could be repaired; the pseudo-label issue is a correctness risk in the mechanism itself. If the FPS cannot recover sub-snippet boundaries from constant-per-snippet targets, the claimed novelty of coarse-to-fine frame-level learning collapses even if the reported numbers survive. The proposed check isolates FPS by comparing against the teacher-only baseline and directly measures within-snippet variation, which is the minimal experiment that would settle the concern. This does not move the reader's conditional verdict; additional experiments and data or code release are still required before acceptance.","tokens_in":11248,"tokens_out":6769,"duration_ms":80433,"concrete_test":"Evaluate a teacher-only baseline on VIEW360: take the trained RTFM snippet teacher used in Section 5.3, duplicate its snippet scores to all 16 frames, and compute frame-level AUC-ROC and AUC-PR under the same protocol as Table 2. Then, for the trained FDPN, restrict the same metrics to frames lying within one snippet of an annotated anomaly boundary and measure the mean absolute deviation of S_frame within each snippet. If the teacher-only baseline matches or exceeds FDPN's 86.00 AUC-ROC, or if FDPN shows no within-snippet score variation and no boundary-restricted improvement over that baseline, the frame-level contribution is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Eq. (4), pseudo-labels P are obtained by thresholding the snippet-level score S_snippet at 0.5 and duplicating the resulting value across all N=16 frames. Consequently, every frame in a positive snippet receives the target 1 (or 0 if the snippet score is below 0.5), and every frame in a negative snippet receives 0. The binary focal loss in Eq. (5) is minimized when S_frame reproduces this constant target, so gradients contain no information about which frames inside a snippet are anomalous. The smoothness loss in Eq. (7) actively penalizes adjacent-frame differences, further pushing predictions toward snippet-level constancy. The frame ranking loss in Eq. (6) selects the top R=48 frames per video, but this is a video-level ranking signal, not a per-snippet boundary signal, and R covers only three snippets. Thus the central claim that FDPN makes better predictions at event boundaries (Figure 2) is not a consequence of the stated training objective. The paper also never ablates the FPS against its own snippet teacher under identical features and masking: Table 4(b) compares saliency versus object detection and Table 5 compares grid sizes, but no row shows the teacher-duplicated frame-level AUC. Without such a baseline, the reported gains on VIEW360, UCF-Crime, and Shanghaitech cannot be attributed to frame-level prediction.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces VIEW360, a new egocentric 360-degree video anomaly detection dataset targeting safety and security situations for visually impaired individuals, and proposes the Frame and Direction Prediction Network (FDPN) for weakly supervised frame-level anomaly detection and anomaly direction classification. FDPN uses a snippet-level teacher network to produce pseudo-labels, a saliency-driven masking preprocessing step, a frame-level prediction subnetwork, and a direction prediction subnetwork. The authors report state-of-the-art AUC-ROC on VIEW360 (86.00), UCF-Crime (88.03), and Shanghaitech (98.51), plus 75.04% direction classification accuracy on VIEW360.","tokens_in":11600,"tokens_out":4988,"duration_ms":56039,"significance":"If validated, the VIEW360 dataset is a useful contribution: it is the first egocentric 360-degree anomaly detection dataset with realistic scenarios and directional labels, and it targets short-duration anomalies that are underrepresented in existing benchmarks. The direction prediction subnetwork and saliency-driven masking are sensible design choices, and the paper includes ablations for masking and direction modules. However, the paper's central claim that FDPN achieves frame-level improvements over snippet-level teachers is not established by the reported experiments: the pseudo-labeling scheme provides no within-snippet supervision, and the comparison against the same teachers used to generate pseudo-labels is not a controlled ablation. Therefore the significance is conditional on additional experiments that isolate the frame-level contribution.","major_comments":[{"comment":"The frame-level training signal does not provide any within-snippet supervision. In Eq. (4), the pseudo-label P is obtained by thresholding the snippet-level score S_snippet at 0.5 and duplicating that single value across all N=16 frames in the snippet, so every frame in a positive snippet receives the same binary target. The binary focal loss in Eq. (5) is then minimized when S_frame reproduces this constant target, and the smoothness loss in Eq. (7) actively penalizes adjacent-frame differences, pushing predictions toward snippet-level constancy. The frame ranking loss in Eq. (6) selects the top R=48 frames per video, which is a video-level ranking signal covering only three snippets, not a per-snippet boundary signal. Consequently, the claimed frame-level boundary improvement illustrated in Figure 2 is not a consequence of the stated training objective. The paper also does not provide an ablation comparing the FPS against its own snippet teacher under identical features and masking: no row in Tables 4 or 5 reports the teacher-duplicated frame-level AUC. Without such a baseline, the reported gains on VIEW360, UCF-Crime, and Shanghaitech cannot be attributed to frame-level prediction.","section":"Section 4.3, Eqs. (4)-(7)"},{"comment":"The comparisons in Tables 2 and 3 are favorable by construction because the same methods used as snippet teachers also appear as baselines. For VIEW360 the teacher is RTFM, for UCF-Crime it is MGFN, and for Shanghaitech it is CLIP-TSA; these exact methods are listed as baselines in the same tables. FDPN is not a stand-alone model but a wrapper that consumes the teacher's snippet scores as pseudo-labels and adds an extra subnetwork, so a comparison against the teacher does not isolate the contribution of frame-level prediction. Moreover, the snippet teacher is selected per dataset after observing dataset characteristics, which raises concerns about selection on the test set. On Shanghaitech the reported gain over CLIP-TSA is only 0.19 AUC-ROC (98.32 vs. 98.51), and on UCF-Crime the gain over MGFN is 1.05 AUC-ROC, with no variance estimates or significance tests. The authors should report repeated runs with fixed random seeds, a teacher-duplicated baseline, and ideally a single teacher configuration across datasets, or justify the per-dataset selection with a validation-based protocol.","section":"Section 5.3, Tables 2 and 3"},{"comment":"There is an inconsistency in the reported feature type for FDPN on Shanghaitech. Table 3 lists FDPN as using I3D features, but Section 5.3 states that for Shanghaitech the snippet network is CLIP-TSA, which operates on CLIP features. Since the FPS concatenates the frame-level features with the snippet-level features F'_snippet produced by the teacher, it is unclear whether the final model uses I3D, CLIP, or both. The authors should clarify which features are used for each dataset and how the CLIP-TSA teacher is integrated into the I3D-based pipeline. This is needed both for reproducibility and for interpreting the Shanghaitech result, where the gain over the teacher is very small.","section":"Section 5.2 and Table 3"}],"minor_comments":[{"comment":"The notation in Eq. (4) is confusing: the condition 'if S+_snippet < 0.5 or S-_snippet' is not a well-formed predicate. It should be rewritten as a clear piecewise definition, for example P = 1 if S+_snippet >= 0.5, and P = 0 if S+_snippet < 0.5 or the video is negative.","section":"Eq. (4)"},{"comment":"Table 5 is difficult to parse because the grid sizes, Top-K values, and AUC numbers are not cleanly separated by rows and columns. Please reformat the table so that each configuration (grid size, Top-K, AUC) is in its own row or column.","section":"Table 5"},{"comment":"The duration-based analysis in Figure 9 reports threshold-dependent improvements but provides no confidence intervals or significance tests. Given that the overall gains are small on some datasets, the authors should indicate how many videos fall into each duration bin and whether the improvement is stable across runs.","section":"Figure 9"},{"comment":"The limitation discussion mentions processing speeds of 1.7 FPS for FDPN and 2.8 FPS for other methods, but does not state the hardware, input resolution, or batching assumptions. Please provide these details so the speed comparison is meaningful.","section":"Section 6, Limitations"}],"recommendation":"major_revision","confidential_remarks":"The strongest part of this submission is the VIEW360 dataset and the direction prediction setup; these are novel and potentially useful for the assistive-technology community. However, the SOTA claim and the frame-level-advantage claim are not yet supported. The authors should be required to add a teacher-duplicated baseline, report results with a fixed or validated snippet-teacher selection, and provide variance estimates. If the frame-level claim cannot be substantiated after these experiments, the paper should be reframed as a dataset-plus-system paper rather than a SOTA frame-level method paper. I also noted that no dataset release URL or code release is mentioned, which limits the practical impact."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe VIEW360 dataset is the thing to take away from this paper. It's the first egocentric 360-degree anomaly detection dataset with direction labels, built for a genuinely under-served user group (visually impaired people), and the collection is thoughtful: IRB approval, privacy blurring, a spread of real locations, and short-duration anomalies that match the stated motivation. The direction prediction task (left-back/center/right-back) is a sensible practical addition. If the dataset is released, it's a usable benchmark.\n\nThe method, FDPN, is a competent mix of known pieces: I3D for snippets, ResNet on saliency-masked frames, PoolFormer for sequence modeling, and multiple-instance ranking losses. That's not a criticism; many good papers are built from known parts. The problem is that the headline claim—frame-level prediction that resolves event boundaries better than snippet-level teachers—doesn't follow from the training objective. Eq. (4) creates pseudo-labels by thresholding the snippet score and duplicating it across all 16 frames. The focal loss then trains the frame subnetwork to reproduce a constant target, and the smoothness loss in Eq. (7) actively pushes adjacent-frame scores closer together. So the network is never given within-snippet supervision. The improvement shown in Figure 2 might happen, but it isn't a consequence of the objective, and the paper never runs the crucial ablation: FPS's output versus its own teacher under identical features and masking. Without that, the gains on UCF-Crime (0.01 over VadCLIP), Shanghaitech (0.16), and VIEW360 could easily be teacher choice or randomness. The per-dataset snippet-network selection (RTFM for VIEW360, MGFN for UCF-Crime, CLIP-TSA for Shanghaitech) confounds the comparison further. No error bars, no code, no data release.\n\nThe authors do state the real-time limitation (1.7 FPS) and the dataset's limited scope, which is honest. The dataset alone is worth publishing. The method needs stronger evidence. I'd send it to peer review with a request for a teacher-vs-frame ablation, released data, and either error bars or a matched comparison. Not a reject; a revise.","headline":"The VIEW360 dataset is a real contribution and the best reason to engage with this paper, but the frame-level SOTA claim is not supported by the training objective or the ablations.","tokens_in":12094,"tokens_out":2512,"would_cite":true,"duration_ms":24845,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"FDPN detects frame-level anomalies in egocentric 360° video and sets state-of-the-art scores on three datasets.","keywords":["video anomaly detection","egocentric 360-degree video","visually impaired assistance","weakly-supervised learning","frame-level prediction","direction prediction","VIEW360 dataset","saliency-driven masking"],"falsifier":"Training FDPN with pseudo-labels randomly shuffled among frames within each snippet, while keeping everything else fixed, would settle the question: if the final frame-level AUC on VIEW360 is unchanged, the subnetwork is only reproducing the teacher's snippet-level segmentation and the claimed boundary improvement does not exist.","tokens_in":11089,"feed_emoji":"🎥","tokens_out":6096,"duration_ms":56295,"temperature":0.7,"pith_summary":"This paper introduces VIEW360, a new egocentric 360-degree video dataset of short, real-world threats such as shoulder surfing and pickpocketing that visually impaired people may face, and proposes FDPN, a network that detects anomalies at frame level and identifies their direction. Existing weakly-supervised methods assign one anomaly score to an entire snippet of frames, which misses brief events; FDPN uses snippet-level predictions as pseudo-labels to train a frame-level subnetwork without extra annotation. On VIEW360 the method reports 86.00 AUC-ROC and 26.97 AUC-PR, outperforming prior methods, and it also reports state-of-the-art frame-level AUC-ROC on UCF-Crime (88.03) and Shanghaitech (98.51). The authors argue this is a first step toward addressing physical safety and security for visually impaired individuals through wearable 360-degree cameras.","feed_headline":"New network catches brief 360° anomalies and says where they are","feed_subtitle":"Frame-level detection on egocentric video sets record AUC on VIEW360, UCF-Crime, and Shanghaitech.","key_machinery":"The core mechanism is the coarse-to-fine pseudo-label pipeline: a pretrained snippet-level anomaly detector (RTFM on VIEW360, MGFN on UCF-Crime, CLIP-TSA on Shanghaitech) provides snippet scores, which are thresholded at 0.5 and duplicated across frames to form pseudo-labels P. The Frame Prediction Subnetwork (FPS), a sequence model inspired by PoolFormer (a pooling-based metaformer) that uses average pooling and 1D convolutions over concatenated ResNet features from saliency-masked frames and I3D snippet features, learns per-frame scores against these pseudo-labels. Saliency-driven image masking, using TASED-Net heatmaps and a top-K grid mask, narrows the input to visually striking regions, and the Direction Prediction Subnetwork (DPS) fuses the same representations with softmax-applied saliency values to output one of three directions. The loss is a weighted sum of binary focal loss, frame ranking loss, smoothness loss, and directional focal loss.","core_discovery":"The central claim is that a coarse-to-fine learning scheme, in which a snippet-level teacher network generates pseudo-labels that are then used to train a frame-level student, yields more precise anomaly detection at event boundaries while also providing directional information. Concretely, FDPN first computes snippet-level anomaly scores from I3D features, thresholds them at 0.5 to create binary pseudo-labels, and replicates these labels across all 16 frames of a snippet. The Frame Prediction Subnetwork, combining masked-image ResNet features and snippet features, is trained with focal, ranking, and smoothness losses to output per-frame scores. A Direction Prediction Subnetwork uses the same features plus saliency heatmaps to classify events as Left back, Center, or Right back. The paper reports state-of-the-art frame-level results on VIEW360 (86.00 AUC-ROC, 26.97 AUC-PR), UCF-Crime (88.03 AUC-ROC), and Shanghaitech (98.51 AUC-ROC), and shows larger gains on short-duration anomalies than a snippet-level baseline.","pith_inferences":["If the gain comes from frame-level refinement rather than the specific teacher, then plugging any stronger snippet-level detector into the coarse-to-fine pipeline should further improve frame-level scores; the paper's per-dataset teacher selection already suggests this dependency.","Because saliency-driven masking operates on a fixed grid in the equirectangular projection, spherical distortion near the poles is left unhandled; a spherical-aware masking scheme could be a natural extension.","The reported processing speed of 1.7 FPS falls short of real-time use, so practical assistive deployment would require a lighter frame-level branch or running the saliency and frame branches only on selected regions; the authors acknowledge this as a limitation."],"forward_implications":["Frame-level anomaly scores enable detection of short, abrupt events such as shoulder surfing and pickpocketing that snippet-level methods miss.","The coarse-to-fine pseudo-label scheme improves performance on existing benchmarks without requiring new frame-level annotations.","Direction predictions (Left back, Center, Right back) provide actionable spatial guidance for a visually impaired user's response.","VIEW360 offers a new benchmark for egocentric 360-degree anomaly detection, with an average anomaly duration of 3.5 seconds."],"supporting_citations":[{"why":"Provides the MGFN snippet network used for UCF-Crime and serves as a comparison baseline for boundary improvement.","marker":"[7]"},{"why":"Provides the RTFM snippet network used for VIEW360 and the pseudo-label thresholding approach; also a comparison baseline.","marker":"[28]"},{"why":"Provides the CLIP-TSA snippet network selected for Shanghaitech.","marker":"[15]"},{"why":"Introduced the UCF-Crime dataset and the smoothness loss used in FPS training.","marker":"[27]"},{"why":"Provided the Shanghaitech dataset used for evaluation.","marker":"[20]"},{"why":"TASED-Net produces the saliency heatmaps used for masking and direction prediction.","marker":"[22]"},{"why":"PoolFormer is the architectural inspiration for the FPS and DPS sequence models.","marker":"[36]"},{"why":"ResNet extracts frame-level features from masked images for FPS and DPS inputs.","marker":"[14]"},{"why":"I3D extracts the snippet-level video features used by the snippet network and FDPN.","marker":"[5]"}],"fun_headline_variants":["360° cam AI flags pickpocketing and pinpoints its direction","Frame-level anomaly spotting from egocentric 360° video","FDPN: detects brief anomalies and tells where they occur","Wearable 360° AI sees all, catches threats frame by frame","New network locates anomalies in 360° egocentric video"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The frame-level subnetwork is trained with pseudo-labels that are identical for every frame in a snippet, so the method assumes these constant labels plus per-frame image features are enough to teach genuinely finer-grained boundaries, an assumption the paper does not directly verify.","fun_headline_variants_meta":{"raw":{"variants":["360° cam AI flags pickpocketing and pinpoints its direction","Frame-level anomaly spotting from egocentric 360° video","FDPN: detects brief anomalies and tells where they occur","Wearable 360° AI sees all, catches threats frame by frame","New network locates anomalies in 360° egocentric video"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000169,"raw_usage":{"total_tokens":1270,"prompt_tokens":957,"completion_tokens":313,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":573,"completion_tokens_details":{"reasoning_tokens":223}},"tokens_in":573,"tokens_out":313,"duration_ms":4860,"temperature":1.0,"reasoning_tokens":223,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:07:34.976195+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Training FDPN with pseudo-labels randomly shuffled among frames within each snippet, while keeping everything else fixed, would settle the question: if the final frame-level AUC on VIEW360 is unchanged, the subnetwork is only reproducing the teacher's snippet-level segmentation and the claimed boundary improvement does not exist.","supporting_citations":[{"cited_title":"Mgfn: Magnitude- contrastive glance-and-focus network for weakly-supervised video anomaly detection","cited_arxiv_id":null,"evidence_quote":"Provides the MGFN snippet network used for UCF-Crime and serves as a comparison baseline for boundary improvement."},{"cited_title":"Weakly-supervised video anomaly detection with robust temporal feature mag- nitude learning","cited_arxiv_id":null,"evidence_quote":"Provides the RTFM snippet network used for VIEW360 and the pseudo-label thresholding approach; also a comparison baseline."},{"cited_title":"Clip-tsa: Clip-assisted temporal self-attention for weakly-supervised video anomaly detec- tion","cited_arxiv_id":null,"evidence_quote":"Provides the CLIP-TSA snippet network selected for Shanghaitech."},{"cited_title":"Real-world anomaly detection in surveillance videos","cited_arxiv_id":null,"evidence_quote":"Introduced the UCF-Crime dataset and the smoothness loss used in FPS training."},{"cited_title":"Fu- ture frame prediction for anomaly detection–a new baseline","cited_arxiv_id":null,"evidence_quote":"Provided the Shanghaitech dataset used for evaluation."},{"cited_title":"Tased-net: Temporally- aggregating spatial encoder-decoder network for video saliency detection","cited_arxiv_id":null,"evidence_quote":"TASED-Net produces the saliency heatmaps used for masking and direction prediction."},{"cited_title":"Metaformer is actually what you need for vision","cited_arxiv_id":null,"evidence_quote":"PoolFormer is the architectural inspiration for the FPS and DPS sequence models."}],"review_version":1}